VACF and vibrational spectrum¶
vacf computes the velocity autocorrelation function and spectrum its Fourier
transform, a vibrational spectral density. Use them as diagnostics of the dynamics.
mliport analyze runs/md vacf --species Li
mliport analyze runs/md spectrum --species Li --taper one-sided-cosine \
--normalization normalized_area
Both require stored velocities and a uniform time axis; velocities are not reconstructed from positions. Both run headlessly and write CSV plus PNG and SVG. Spectrum area normalization supports both NumPy 1.x and 2.x.
Options you will use¶
| Option | Meaning |
|---|---|
--species SPECIES |
the species to correlate (all atoms when omitted) |
--method fft\|direct |
the correlation algorithm (default fft) |
--taper one-sided-cosine\|none |
window applied before the transform (spectrum; default one-sided-cosine) |
--normalization normalized_area\|raw_spectrum |
spectrum normalization (spectrum; default normalized_area) |
--start-frame, --stop-frame, --include-equilibration |
the analysis window |
--positions-convention, --frame-interval-fs |
metadata for external trajectories |
What to watch¶
The saved-frame interval sets the Nyquist frequency, not the integration timestep. If it is too coarse for the spectral range you asked for, mliport reports insufficient sampling instead of silently aliasing.
Spectra depend on the frame spacing, the taper and the trajectory length, so report all three. A spectrum from a classical MLIP trajectory is not a quantitative phonon calculation and carries no quantum nuclear effects. Compare spectra only inside one model and timestep protocol. A short trajectory gives a noisy but finite spectrum; do not read its frequencies as converged vibrational modes.
mliport analyze runs/md-300K spectrum --species O --method fft \
--taper one-sided-cosine --normalization normalized_area
The acceptance suite runs the plain forms on the 1 ns GRACE trajectory: